Touch of Privacy: A Homomorphic Encryption-Powered Deep Learning Framework for Fingerprint Authentication¶
Authors: U. Sumalatha; K. Krishna Prakasha; Srikanth Prabhu; Vinod C. Nayak
Year: 2025
Venue: IEEE Access
type: application collection: ppml status: auto
Links: DOI
Summary¶
Auto-extracted summary (unreviewed)
Builds a privacy-preserving fingerprint authentication system in which a CNN extracts fingerprint features that are encrypted with CKKS via TenSEAL, and matching is performed by Euclidean similarity in the encrypted domain so the server never sees raw biometrics; decryption happens only on the client. On a 90-user, 7,200-image subset of SOCOFing the system reaches 99.06% test accuracy, EER 0.40% and FAR 0%, with an encrypted comparison in 0.025 s and 0.136 s total per fingerprint despite a 157.26x encryption overhead.
Abstract¶
Deep learning and fully homomorphic encryption (FHE) are integrated for privacy-preserving fingerprint recognition. Convolutional neural network (CNN) extract fingerprint features encrypted using the Cheon-Kim-Kim-Song (CKKS) FHE scheme. TenSEAL ensures all computations occur in the encrypted domain, preventing raw biometric data exposure during authentication. A subset of the SOCOFing dataset, comprising 7,200 altered fingerprints from 90 individuals across three difficulty levels, is used for training (80%), validation (10%), and testing (10%). One real fingerprint per user is encrypted and stored for authentication, reducing computational complexity. The CNN classifies encrypted features without decryption, ensuring secure authentication. The system, utilizing Euclidean similarity, achieves 99.06% test accuracy with a loss of 0.1692, a True Acceptance Rate (TAR) of 99.19%, a False Rejection Rate (FRR) of 0.81%, a False Acceptance Rate (FAR) of 0%, and a True Rejection Rate (TRR) of 100%, with a minimal Equal Error Rate (EER) of 0.40%. Encrypted templates require only 31 MB for 90 users, averaging 344 KB per fingerprint per user. Despite a 157.26× encryption overhead, remains feasible for real-time applications, completing an encrypted comparison in 0.025 seconds with a total processing time of 0.136 seconds per fingerprint. Adhering to ISO/IEC 24745 biometric protection standards, the system ensures irreversibility, unlinkability, and renewability of biometric templates. Decryption occurs only on the client side, keeping raw fingerprint data inaccessible to the server. The results confirm the feasibility of FHE for secure fingerprint authentication in cloud-based environments, benefiting applications in healthcare, banking, and access control.
Source: pdf
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | CNN feature extraction combined with CKKS-encrypted template matching for fingerprint authentication | llm | 0.75 |
| contribution | Threshold-based verification using Euclidean similarity on encrypted features | llm | 0.75 |
| contribution | One encrypted real fingerprint per user stored to reduce computational cost | llm | 0.75 |
| contribution | Compliance with ISO/IEC 24745 (irreversibility, unlinkability, renewability) | llm | 0.75 |
| contribution | Evaluation of accuracy, biometric error rates, storage and latency on SOCOFing | llm | 0.75 |
| schemes | CKKS | llm | 0.75 |
| libraries | TenSEAL | llm | 0.75 |
| models | CNN | llm | 0.75 |
| datasets | SOCOFing (subset: 7,200 altered fingerprints, 90 users) | llm | 0.75 |
| tasks | fingerprint authentication | llm | 0.75 |
| tasks | biometric recognition | llm | 0.75 |
| tasks | secure cloud-based biometric matching | llm | 0.75 |
| threat model | client-server; decryption only on client, server never accesses raw fingerprint data | llm | 0.75 |
| interaction | non-interactive FHE | llm | 0.75 |
| limitations | 157.26x encryption overhead relative to plaintext processing | llm | 0.75 |
| limitations | Evaluated on a 90-user subset of SOCOFing | llm | 0.75 |
| llm relevance | low | llm | 0.75 |
| key result | test accuracy: 99.06% (SOCOFing subset, Euclidean similarity) | llm | 0.75 |
| key result | TAR / FRR / FAR / TRR: 99.19% / 0.81% / 0% / 100% (SOCOFing subset) | llm | 0.75 |
| key result | EER: 0.40% (SOCOFing subset) | llm | 0.75 |
| key result | encrypted comparison latency: 0.025 s (0.136 s total per fingerprint) (CKKS via TenSEAL) | llm | 0.75 |
| key result | encryption overhead: 157.26x (vs plaintext) | llm | 0.75 |
| key result | template storage: 31 MB for 90 users (344 KB per fingerprint) (encrypted templates) | llm | 0.75 |
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: CKKS 10 [p.1], Multi-key HE 3 [p.4], Leveled / somewhat HE 3 [p.2]
- library: TenSEAL 18 [p.1], Microsoft SEAL 6 [p.2], Concrete (Zama) 3 [p.3], Concrete ML 2 [p.3], TensorFlow 1 [p.9], ONNX 1 [p.3]
- model: CNN 30 [p.1], LeNet 1 [p.3], k-means 1 [p.4]
- dataset: Fingerprint (FVC) 133 [p.1]
- hardware: GPU 3 [p.8], NVIDIA A100 1 [p.8], NVIDIA Jetson 1 [p.4], Intel Xeon 1 [p.9], TPU 1 [p.16], CUDA 1 [p.9]
- technique: Convolution 9 [p.1], Quantization 3 [p.3], Bootstrapping 2 [p.2], Softmax 2 [p.10], Rotations 2 [p.3], Fine-tuning 2 [p.10]
- assumption: LWE 3 [p.14]
- task: Medical / healthcare 9 [p.1], Finance 3 [p.1], Clustering 2 [p.4], IoT / edge 2 [p.4]
- topic: Homomorphic encryption 59 [p.1], Privacy-preserving machine learning 5 [p.3], Private / secure inference 2 [p.3]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/ppml/applications/sumalatha-2025_touch-of-privacy-he-powered-deep-learning-for-fingerprint-authentication.pdf |
ppml/applications | 17 | 1.5 MB | 2a3055d7e35c |
was Sumalatha et al. - 2025 - Touch of Privacy A Homomorphic Encryption-Powered Deep Learning Framework for Fingerprint Authentic.pdf |
Database key: touch-of-privacy-a-homomorphic-encryption-powered-deep-learning · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set touch-of-privacy-a-homomorphic-encryption-powered-deep-learning COLUMN VALUE